Vault Structure

In late March I killed a custom Electron app I had been building called TruPath OS. It was supposed to be the operating system for my holding company — typed schemas, custom UI, a 20-agent build team scoped to ship it. I wrote about that postmortem in Issue #0001.

The replacement was an Obsidian vault. Just markdown files in numbered folders, opened in Obsidian, edited by me and read by Claude Code. Today, six weeks in, it runs every operational and strategic concern across Quantum Caddy, Mile High Golf, and Parley. This bundle is the template.

Why a vault, not a SaaS stack

The reasonable default for most operators is a stack: Notion for docs, Asana or Linear for tasks, Google Drive for files, Slack for chat, Calendar + Email + GitHub + Figma. For a 30-person company with funded operations, this is fine. For a solo operator running multiple ventures with AI agents, it's not.

The reason isn't that any individual tool is bad. The reason is that AI agents cannot effectively read across SaaS boundaries. Your agent can search Notion, query Asana, browse Drive, summarize Gmail — but it can't easily *cross-reference* them. The decision we made in Tuesday's meeting lives in three places: the meeting notes in Notion, the action item in Asana, the email confirming it in Gmail. Each is partially correct. None of them is queryable as a whole. The cost of context-stitching falls on you.

Move everything you reasonably can into a single Obsidian vault — markdown files, structured frontmatter, in one git repo — and several things change at once.

Cross-references become free. A meeting note links to a decision, which links to a person, which links to a venture, which links to a wiki page. Obsidian's graph view shows you the actual structure of your work.

Agents read everything in one pass. "What did Sarah say about the lease last month" is a single grep. Three files. Same answer you'd get from a 15-minute archaeology run.

Versioning is git, not vendor history. You own your data.

Your agent's context window holds your work, not tooling chrome. Markdown is the most token-efficient format an agent will ever read.

The 13 folders

00-Inbox/ — raw captures, processed weekly.

01-Labs/ — content engine, newsletter, playbook.

02-Holding-Company/ — parent entity, finance, legal, IR, strategy.

03-Venture-A/ — first venture (rename to your slug).

04-Venture-B/ — second venture.

05-Agents/ — agent definitions, rosters, activation guides.

06-People/ — contact directory.

07-Decisions/ — ADRs, decision log.

08-Meetings/ — meeting notes.

09-Research/ — Karpathy-style synthesis wiki.

10-Session-Logs/ — Claude Code session summaries.

11-Archive/ — deprecated patterns.

12-Templates/ — the 13 templates.

13-Social-Media/ — brand voice and platform-specific drafts.

The numbering is deliberate. Without numbers, "Strategy" and "Stuff" sit next to each other. With numbers, the order *is* your operational priority.

One folder per concern, never two. Decisions live in 07-Decisions/. Always.

Why the wiki layer matters

Inside 09-Research/ is the synthesis surface. Raw chat history, raw research, raw brainstorming all flow into 09-Research/Inbox/. The synthesis workflow compresses them into focused, cross-linked wiki pages with one-line summaries.

For QC, this means the work I did researching cornhole physics, sensor selection, ACL competition rules, and venue installation requirements lives in 12 wiki pages with bidirectional links — not 200 chat transcripts. When I ask Claude a question about the QC sensor pivot from FSR to radar, it reads three wiki pages, not three months of conversation history.

This is the single most expensive piece of operating discipline to build from scratch. The bundle ships with the workflow, the page format, the index pattern, and a worked example.

The agent layer

The vault includes a 5-agent placeholder model: a Chief of Staff, two Venture Directors, an Engineering specialist, and a Legal/IP specialist. Replace the names with the agents your portfolio actually needs.

The pattern that's worth keeping is the *number*. After eight months of running an AI-augmented operation, my honest take is that a small set of named, scoped specialists is more reliable than a sprawling cast of agents that don't know their lane. I started with 14 agents. I cut to 5. The work got better.

What changes over time

Three phases for most adopters.

Weeks 1-2: friction. The structure feels rigid. You're constantly looking up which folder is for what.

Weeks 3-8: convergence. You stop thinking about where things go. Agents start reading across your work in ways that feel like magic the first time.

Months 3+: compounding. The wiki has accumulated enough synthesis that knowledge questions are fast. New patterns absorb into the structure without much disruption.

After about six months, the structure feels less like a tool and more like a shape that your work organically takes.

What's in the bundle

A working template vault — 13 folders, all 13 templates (anonymized), an example wiki page, an example agent file, a CLAUDE.md template — plus a getting-started guide that walks through the first 30 minutes, the framing essay, and an MIT license.

— Michael, from the lab